Import Libraries

import numpy as np
import pandas as pd
import re
import datetime

from plotly.offline import init_notebook_mode, iplot, plot
import plotly as py
import plotly.express as px
init_notebook_mode(connected=True)
from plotly.subplots import make_subplots
import plotly.graph_objs as go


import matplotlib.pyplot as plt

import os

Load Data

air_visits = pd.read_csv('./Data/air_visit_data.csv')
air_reserve = pd.read_csv('./Data/air_reserve.csv')
hpg_reserve = pd.read_csv('./Data/hpg_reserve.csv')
air_store = pd.read_csv('./Data/air_store_info.csv')
hpg_store = pd.read_csv('./Data/hpg_store_info.csv')
holidays = pd.read_csv('./Data/date_info.csv')
store_ids = pd.read_csv('./Data/store_id_relation.csv')
test_data = pd.read_csv('./Data/sample_submission.csv')

Quick Glance at File Summaries

# insert whichever dataframe you want to view here:
air_visits.describe()

	visitors
count	252108.000000
mean	20.973761
std	16.757007
min	1.000000
25%	9.000000
50%	17.000000
75%	29.000000
max	877.000000

# quick look at number of rows and cols
air_visits.shape

(252108, 3)

Check for Missing Values

# can apply this to all files
missing_values = air_visits.isna().sum()
missing_values

air_store_id    0
visit_date      0
visitors        0
dtype: int64

Individual Feature Visualizations

# we will take a look at the number of visist to the AIR restaurants

# first plot wll look to group visits by visit date and sum the # of visitors
p1 = air_visits.groupby(['visit_date'], as_index=False).sum()
fig = px.line(p1, x='visit_date', y='visitors', title='Total Visitors by Visit Date')
fig.show()

# display the value (in chart) as log based visitors, but have the tick marks are the "true" value
fig=px.histogram(air_visits, x='visitors', title='Occurences of Visitor count')
fig.update_xaxes(type='log', range=[0,2])
fig.update_layout(bargap=0.05, xaxis={'tickmode':'array', 
    'tickvals': pd.to_numeric([f"{n:.1g}" for n in np.geomspace(1, air_visits['visitors'].max(), 15)]),})
fig.show()

# plot the max visitors by air_id and the count of occurences
p1 = air_visits.groupby(['air_store_id'], as_index=False).max()
fig=px.histogram(p1, x='visitors', title='Occurences of Max Visitor count by store id')
fig.update_layout(bargap=0.1)
fig.show()

# Creating a temp dataframe copy from air_visits in order to generate a day of the week column.
# Day of week: Monday = 0 Sunday = 6
df_tmp = air_visits.copy()
df_tmp['date'] = pd.to_datetime(df_tmp['visit_date'])
df_tmp['wday'] = df_tmp['date'].dt.dayofweek.astype(str)

p1 = df_tmp.groupby(['wday'], as_index=False).median()
fig = px.bar(p1, x='wday', y='visitors', color = 'wday')
fig.update_layout(coloraxis={'showscale': False}, showlegend=False, xaxis={'tickmode': 'array', 'tickvals': [0,1,2,3,4,5,6], 'ticktext': ['Mon', 'Tue', 'Wed', 'Thur', 'Fri', 'Sat', 'Sun']},
    xaxis_title='Day of the Week', yaxis_title= 'Median Visitors')
fig.show()

# Extractingt he month from the temp dataframe used above and creatinga new columns called 'Month' in order to graph the median visitors by month
df_tmp['month'] = df_tmp['date'].dt.month.astype(str)

p2 = df_tmp.groupby(['month'], as_index=False).median()
fig = px.bar(p2, x='month', y='visitors', color = 'month')
fig.update_layout(coloraxis={'showscale': False}, showlegend=False, xaxis={'tickmode': 'array', 'tickvals': [1,2,3,4,5,6,7,8,9,10,11,12], 
    'ticktext': ['Jan', 'Feb', 'Mar', 'Apr', 'May', 'Jun', 'Jul', 'Aug', 'Sep', 'Oct', 'Nov', 'Dec']},
    xaxis_title='Month', yaxis_title= 'Median Visitors')
fig.show()

# plotting the above except by month and year
df_tmp['year'] = df_tmp['date'].dt.year.astype(str)
df_tmp['year_month'] = df_tmp['year'] + '-' + df_tmp['month']

p3 = df_tmp.groupby(['year_month'], as_index=False).agg({'visitors': 'median'})
p3['month'] = pd.to_datetime(p3['year_month']).dt.month.astype(str)

fig = px.bar(p3, x='year_month', y='visitors', color='month')
fig.update_layout(coloraxis={'showscale': False}, showlegend=False, xaxis_title='Year-Month', yaxis_title= 'Median Visitors')
fig.show()

Findings:

Plot 1: Visitors over time

• The visitor count shows a step structure (each year bumps up the "average" of the cycle) which may indicate additional restaurants being added to the Air database
• Overall there is a clear weekly, I am assuming as we cna investigate further, cycle to the visitor cadence that is apparent in all the months provided

Plot 2 & 3: Occurences of Visitor Count // Occurences of Max Visitor Count by Store ID

• Visitor counts seem to mostly converge around the 6-0 visitors per day across all days recorded
• Grouping each sotre id's data together, the highest visitor count that occurs the most frequent is 30-39 visitors with 105 occurences across the stores, some peaks extend to 100+ people, will have to check those isolated incidents of extreme visitor counts

Plot 4: Median Visitors by Day of Week

• Friday through Sunday appear to be the more popular days which a drop off on Mondays and Tuesdays. Saturday has the peak visitation

Plot 5 & 6: Median Visitors by Month // Median Visitors by Month-Year

• This shows the median across each year, which holds fairly consistent for common months accross the dataset, within the year there is variation with December being the most popular month, and the prior month November being on the lower side of the medians listed.
• Mar, April, and May show as above average as well (not as much as December) but remain very busy months.

As we will be forecasting the last week of April into May of 2017, we can plot the corresponding timeframe from the 2016 training data provided. We'll extend the timeframe a little to have two months worth of data (Apr 15-Jun 15 of 2016)

df_apr_may_2016 = df_tmp[ (df_tmp['date'] > '2016-04-15') & (df_tmp['date'] < '2016-06-15') ]
p1 = df_apr_may_2016.groupby(['date'], as_index=False).agg({'visitors':'sum'})

fig1 = px.line(p1, x='date', y='visitors')
fig2 = px.scatter(p1, x='date', y='visitors', trendline='rolling', trendline_options={'window': 3}, color_discrete_sequence=px.colors.qualitative.Pastel1)
fig3 = go.Figure(data=fig1.data + fig2.data)
fig3.update_layout(xaxis_title='Date', yaxis_title= 'All Visitors')
fig3.show()

# fig=go.Figure()
# fig.add_trace(go.Scatter(x=p1['date'], y=p1['visitors'], line_shape='spline'))
# fig.add_trace(go.Scatter(x=p1['date'], y=p1['visitors'], mode='lines'))
# fig.show()

Looking at this plot, we can see the blue line which corresponds to the raw data (total visitors by day), and an approximated pink line with a smoothing fit. There is a drop compared to the overall pattern from May 1st to just before May 8th, this corresponds to the holiday Golden Week.

Now we can switch gears from the visitor data (true visits) to the reservation data, we will continue with the AIR data.

We can do a similar datetime visualization with the reservtion dataset as this includes a field that shows the visitor count on that same day. We can also take a look at the time difference bewtween the time the reservation was made at, and when the reservation was made for.

# working with the air_reservation data now, will work to extract month, hour, and weekday values for each of the reservation fields and the visit fields
foo = air_reserve.copy()

# converting and extracting the reservation data
foo['reserve_date'] = pd.to_datetime(foo['reserve_datetime'])
foo['reserve_hour'] = foo['reserve_date'].dt.hour
foo['reserve_wday'] = foo['reserve_date'].dt.dayofweek
foo['reserve_date'] = foo['reserve_date'].dt.date

# converting and extracting the visit data
foo['visit_date'] = pd.to_datetime(foo['visit_datetime'])
foo['visit_hour'] = foo['visit_date'].dt.hour
foo['visit_wday'] = foo['visit_date'].dt.dayofweek
foo['visit_date'] = foo['visit_date'].dt.date

# calculating the difference in time from reservation to visit (visit time - reservation time, result in HOURS)
foo['diff_hour'] = pd.to_datetime(foo['visit_datetime']) - pd.to_datetime(foo['reserve_datetime'])
foo['diff_hour'] = foo['diff_hour']/np.timedelta64(1, 'h')

# calculating the difference in time from reservation to visit (visit time - reservation time, result in DAYS)
foo['diff_day'] = pd.to_datetime(foo['visit_datetime']) - pd.to_datetime(foo['reserve_datetime'])
foo['diff_day'] = foo['diff_day'].dt.days

p1 = foo.groupby(['visit_date'], as_index=False).agg({'reserve_visitors':'sum'})
fig = px.line(p1, x='visit_date', y='reserve_visitors')
fig.update_xaxes(title='AIR Visit Date')
fig.update_yaxes(title='Total Reservations')
fig.show()

p2 = foo.groupby(['visit_hour'], as_index=False).agg({'reserve_visitors': 'sum'})
fig2 = px.bar(p2, x='visit_hour', y='reserve_visitors')
fig2.update_xaxes(title='AIR Visit Hour')
fig2.update_yaxes(title='Total Reservations')
fig2.show()

# only showing time gaps shorter than 5 days out from the time of reservation as some values go out to over a year
p3 = foo.groupby(['diff_hour'], as_index=False).agg({'reserve_visitors': 'sum'})
p3 = p3[ p3['diff_hour'] < 25*5 ]

fig3 = px.bar(p3, x='diff_hour', y='reserve_visitors')
fig3.update_xaxes(title='AIR: Time from Reservation to Visit (hrs)')
fig3.update_yaxes(title='Total Reservations')
fig3.show()

Findings:

Plot 1: Total Reservations by Date

• We can see that 2016 had much fewer reservations compared to 2017, with a segment of time where there were no reservations made at all
• The tapering seen at the end of 2017 is likely due to this time frame being at the end of the training time frame provided ##### Plot 2: Total Reservations by Hour in the day
• The most frequented time to set reservations for was 6 and 7 pm, with evenings overall being the favorited time for reservations ##### Plot 3: Total Reservations - By Time from Reservation to Visit (hours)
• The majority of reservations are made for same day visits, with the peaksbeing within a hour of scheduling
• In combination with the prior findingt hat evenings was the most popular time frame, we can see that reservtions follow a pattern if not for the same day, then it will be scheduled in the evening for a following evening
• This plot was truncated as there are a few reservations that can go up to a year ahead of the planned visit, we can take a look at these lines below:

tmp = foo.sort_values(['diff_day'], ascending=False)
tmp.iloc[:5, [0,1,2,11]]

	air_store_id	visit_datetime	reserve_datetime	diff_day
59263	air_e7fbee4e3cfe65c5	2017-02-07 20:00:00	2016-01-11 17:00:00	393
59257	air_e7fbee4e3cfe65c5	2017-02-07 20:00:00	2016-01-12 20:00:00	392
59262	air_e7fbee4e3cfe65c5	2017-02-07 20:00:00	2016-01-18 18:00:00	386
50826	air_2a485b92210c98b5	2017-01-18 20:00:00	2016-01-02 00:00:00	382
50805	air_2a485b92210c98b5	2017-01-18 20:00:00	2016-01-02 20:00:00	382

This is the top 5 longest differences (by day) in reservations to visit dates. This contains only 2 unique air store ids, must be very indemand locations where this type of reservation is standard or they could be input errors with incorrect years.

We will now shift gears towards looking at the HPG data, similar dataset as the AIR info but for a different app to schedule reservations. We will perform the same chartings as above for:

• Total reservations by date
• Total reservations by visit hour
• Total reservations by time from reservations to visit (in hours)

# working with the hpg_reservation data now, will work to extract month, hour, and weekday values for each of the reservation fields and the visit fields
foo = hpg_reserve.copy()

# converting and extracting the reservation data
foo['reserve_date'] = pd.to_datetime(foo['reserve_datetime'])
foo['reserve_hour'] = foo['reserve_date'].dt.hour
foo['reserve_wday'] = foo['reserve_date'].dt.dayofweek
foo['reserve_date'] = foo['reserve_date'].dt.date

# converting and extracting the visit data
foo['visit_date'] = pd.to_datetime(foo['visit_datetime'])
foo['visit_hour'] = foo['visit_date'].dt.hour
foo['visit_wday'] = foo['visit_date'].dt.dayofweek
foo['visit_date'] = foo['visit_date'].dt.date

# calculating the difference in time from reservation to visit (visit time - reservation time, result in HOURS)
foo['diff_hour'] = pd.to_datetime(foo['visit_datetime']) - pd.to_datetime(foo['reserve_datetime'])
foo['diff_hour'] = foo['diff_hour']/np.timedelta64(1, 'h')

# calculating the difference in time from reservation to visit (visit time - reservation time, result in DAYS)
foo['diff_day'] = pd.to_datetime(foo['visit_datetime']) - pd.to_datetime(foo['reserve_datetime'])
foo['diff_day'] = foo['diff_day'].dt.days

p1 = foo.groupby(['visit_date'], as_index=False).agg({'reserve_visitors':'sum'})
fig = px.line(p1, x='visit_date', y='reserve_visitors')
fig.update_xaxes(title='HPG Visit Date')
fig.update_yaxes(title='Total Reservations')
fig.show()

p2 = foo.groupby(['visit_hour'], as_index=False).agg({'reserve_visitors': 'sum'})
fig2 = px.bar(p2, x='visit_hour', y='reserve_visitors')
fig2.update_xaxes(title='HPG Visit Hour')
fig2.update_yaxes(title='Total Reservations')
fig2.show()

# only showing time gaps shorter than 5 days out from the time of reservation as some values go out to over a year
p3 = foo.groupby(['diff_hour'], as_index=False).agg({'reserve_visitors': 'sum'})
p3 = p3[ p3['diff_hour'] < 25*5 ]

fig3 = px.bar(p3, x='diff_hour', y='reserve_visitors')
fig3.update_xaxes(title='HPG: Time from Reservation to Visit (hrs)')
fig3.update_yaxes(title='Total Reservations')
fig3.show()

Findings:

As these are the same plots, I will only note the differences in general compared to the AIR data versions

• As this dataset is substantially larger, we can see a much more defined pattern in the reservations made over time with the spike in December of 2016 to be a clear 'out of the norm' situation, likely due to holidays.
• Same as before, the most common reservations are made for evening hours
• Different than the AIR data, the HPG data shows a slight cange tot he cycle of time between reservations and visit. We see that the first few hours after reservations is not the peak of the data, rather the 48 and even 72 hour points show a higher volume of reservations

Let's take a look at a heatmap indicating where these restaurant locations are in Japan

# using air_store we can take a look at the lats and longs provided of the stores
fig = px.density_mapbox(air_store, lat='latitude', lon='longitude', mapbox_style='carto-positron')
fig.update_layout()
fig.show()

To get a more numerical view, lets plot out the count of restaurants by area, and count of categories. we'll stick to the top 15 if the lists become to cumbersome to view.

foo = air_store.copy()

p1 = foo.groupby(['air_genre_name'], as_index=False).size()
p1.rename(columns={'size':'count'}, inplace=True)
p1.sort_values(['count'], ascending=False, inplace=True)

fig = px.bar(p1, x='count', y='air_genre_name', orientation='h', color='air_genre_name')
fig.update_xaxes(title='Count of Genre Type (AIR)')
fig.update_yaxes(title='Genre Name')
fig.update_layout(showlegend=False)
fig.show()

p2 = foo.groupby(['air_area_name'], as_index=False).size()
p2.rename(columns={'size':'count'}, inplace=True)
p2.sort_values(['count'], ascending=False, inplace=True)
p2 = p2.iloc[:15, :]

fig2 = px.bar(p2, x='count', y='air_area_name', orientation='h', color='air_area_name')
fig2.update_xaxes(title='Num of Restaurants (AIR)')
fig2.update_yaxes(title='Area Name (Top 15)')
fig2.update_layout(showlegend=False)
fig2.show()

Findings:

• In general, we can see that the top genre is the Izakaya (informal Japanese bar, typical location for after-work drinking, similar to a pub), which is followed closely by Cafes.
• This dataset has very little restaurants listed as Karaoke/International/Asian, with "Other" being listed in the middle of the pack and lower on the list something called "Creative Cuisine"
• In terms of the areas, Fukuoka has the most number of restaurants followed by 5 Tokyo areas which Shibuya being the highest among this subset

Let's plot smilar charts using the HPG data and see if there are any key differences or, like the previous plot comparison, if any trends solidify due to the larger volume of data.

# using hpg_store we can take a look at the lats and longs provided of the stores
fig = px.density_mapbox(hpg_store, lat='latitude', lon='longitude', mapbox_style='carto-positron')
fig.update_layout()
fig.show()

foo = hpg_store.copy()

p1 = foo.groupby(['hpg_genre_name'], as_index=False).size()
p1.rename(columns={'size':'count'}, inplace=True)
p1.sort_values(['count'], ascending=False, inplace=True)
p1 = p1.iloc[:15, :]

fig = px.bar(p1, x='count', y='hpg_genre_name', orientation='h', color='hpg_genre_name')
fig.update_xaxes(title='Count of Genre Type (HPG)')
fig.update_yaxes(title='Genre Name (Top 15)')
fig.update_layout(showlegend=False)
fig.show()

p2 = foo.groupby(['hpg_area_name'], as_index=False).size()
p2.rename(columns={'size':'count'}, inplace=True)
p2.sort_values(['count'], ascending=False, inplace=True)
p2 = p2.iloc[:15, :]

fig2 = px.bar(p2, x='count', y='hpg_area_name', orientation='h', color='hpg_area_name')
fig2.update_xaxes(title='Num of Restaurants (HPG)')
fig2.update_yaxes(title='Area Name (Top 15)')
fig2.update_layout(showlegend=False)
fig2.show()

Findings:

• For the HPG data, it looks like the top genre is a more broad category of "Japenese style" with "International Cuisine" following behind. This dataset contained a larger variety so I have limited this to the top 15 genres
• We find that Tokyo is again at the top in terms of number of restaurants, to view this in more general terms we could truncate the provided area names which appear to specify a more precise location as you progress through the provided format
• Interestingly there is a None category allocated to restaurants and this appears to encompass over 100 store ids #### Let's try to trim down the area names and get more broad numbers

# air restaurants
foo = air_store.copy()

foo['trunc_area'] = foo['air_area_name'].str.split(r'[-\s]', expand=True)[0]

p1 = foo.groupby(['trunc_area'], as_index=False).size()
p1.rename(columns={'size':'count'}, inplace=True)
p1.sort_values(['count'], ascending=False, inplace=True)

fig = px.bar(p1, x='count', y='trunc_area', orientation='h', color='trunc_area')
fig.update_xaxes(title='Num of Restaurants (AIR)')
fig.update_yaxes(title='Prefecture Name')
fig.update_layout(showlegend=False)
fig.show()

# hpg restaurants
foo = hpg_store.copy()

foo['trunc_area'] = foo['hpg_area_name'].str.split(r'[-\s]', expand=True)[0]

p1 = foo.groupby(['trunc_area'], as_index=False).size()
p1.rename(columns={'size':'count'}, inplace=True)
p1.sort_values(['count'], ascending=False, inplace=True)

fig = px.bar(p1, x='count', y='trunc_area', orientation='h', color='trunc_area')
fig.update_xaxes(title='Num of Restaurants (HPG)')
fig.update_yaxes(title='Prefecture Name')
fig.update_layout(showlegend=False)
fig.show()

Taking a look at the adjusted plots taking into consideration just the top level prefecture, Tokyo is by far the most accounted for area with Osaka and Fukuoka filling out the top 3 for number of restaurants.

We now shift focus to taking a top level look at the holidays in Japan from the provided holidays dataset. This will provide some context to the total number of holidays and their potential impact on reservations/visitors. We can also look at where these holidays land along our date range for the provided data.

# looking at the date_info file imported as holidays
p1 = holidays.groupby('holiday_flg', as_index=False).size()

fig = px.bar(p1, x='holiday_flg', y='size', color='holiday_flg')
fig.update_xaxes(title='Holiday Flag')
fig.update_yaxes(title='Count')
fig.update_layout(showlegend=False, coloraxis={'showscale': False})

# prepare date split for 2016 and 2017 data
p2 = holidays.copy()
p2 = p2.set_index(['calendar_date'])
p2 = p2.loc['2016-04-15' : '2016-06-01']
p2.reset_index(inplace=True)

p3 = holidays.copy()
p3 = p3.set_index(['calendar_date'])
p3 = p3.loc['2017-04-15' : '2017-06-01']
p3.reset_index(inplace=True)

# make dual plot with 2016 and 2017 
fig = make_subplots(rows=2, cols=1, vertical_spacing=0.1)

fig.add_trace(go.Scatter(x=p2['calendar_date'], y=p2['holiday_flg'], mode='markers', marker_color=p2['holiday_flg']), row=1, col=1)
fig.add_trace(go.Scatter(x=p3['calendar_date'], y=p3['holiday_flg'], mode='markers', marker_color=p3['holiday_flg']), row=2, col=1)

fig.update_layout(height=600, width=800, title_text="Holidays in dataset", showlegend=False, title_x=0.5)
fig.show()

holidays['holiday_flg'].agg({'holiday_flg': 'mean'})

holiday_flg    0.067698
Name: holiday_flg, dtype: float64

We find that the same days were holidays in 2016 and 2017 in late April and May, additionally, almost 7% of the dates listed are holidays

# create unique date set for the training data
foo = air_visits.copy()
foo.rename(columns={'visit_date':'date'}, inplace=True)
foo.drop_duplicates(subset=['date'], inplace=True)
foo['dset'] = 'train'

# create unique date set for the testing data
bar = test_data.copy()
bar[['foo', 'bar', 'date']] = bar['id'].str.split('_', -1, expand=True)
bar.drop_duplicates(subset=['date'])
bar['dset'] = 'test'

baz = pd.concat([foo, bar], axis=0, ignore_index=True)
baz['year'] = pd.DatetimeIndex(baz['date']).year
baz['year'] = baz['year'].astype(str)
baz['date'] = baz['date'].str.replace('2016', '2017', regex=False)
baz = baz[baz['date'].notna()]

fig = px.scatter(baz, x='date', y='year', color='dset')
fig.update_traces(marker={'size':10 })
fig.layout.xaxis.tickvals = pd.date_range('2017-01', '2018-01', freq='MS')
fig.layout.xaxis.tickformat = '%b'
fig.show()

We can see that the test dataset, in red, comes in about the last week of April and encompasses up to June of 2017. Our training set has a full year+ of data to work with.

We can start to combine features and analyze their relationships, which in trn need to be interpreted in the context of the individual features distributions, similar to what we did above for some features.

We'll start with average AIR restaurant visitors in relation to the type of cuisine (genre)

# join our visit information and genre information into a new df
foo = air_visits.merge(air_store, on = 'air_store_id', how='left')
foo = foo.groupby(['visit_date', 'air_genre_name'], as_index=False).agg({'visitors': 'mean'})

genres = list(foo.groupby('air_genre_name'))
subplot_titles = [genre[0] for genre in genres]

rows = 4
cols = 4

fig = make_subplots(rows=rows, cols=cols, vertical_spacing=0.15, subplot_titles=subplot_titles, shared_yaxes=True)

for i, l in enumerate(genres):
    row = i // cols + 1
    col = (i % rows) + 1
    df = l[1]
    fig.add_trace(go.Scatter(x=df['visit_date'], y=df['visitors'], showlegend=False), row=row, col=col)

fig.update_layout(height=800, width=1300, title_text="Average number of visitors by genre (AIR)")

# update the yaxis values for each plot
for ax in fig['layout']:
    if ax[:5] == 'yaxis':
        fig['layout'][ax]['range'] = [0,100]

fig.show()

Findings:

• Most of the visitor values appear to fall between 10 and 100 but even further it seems most have means around 25 visitors daily
• We previously saw that the genre's Karaoke, International Cuisine, and Asian were lower in total count, and it is reflected here with plots that appear more volatile with the Asian genre still showing a higher average than the others in this subset

We can see from these plots that there is some semplance of a weekly pattern which we saw in earlier plots of visitor data, with this we can take a look at the mean values by day of week and genre.

foo = air_visits.merge(air_store, on = 'air_store_id', how='left')
foo['wday'] = pd.to_datetime(foo['visit_date']).dt.dayofweek
foo['wday'] = foo['wday'].astype(str)
p1 = foo.groupby(['wday', 'air_genre_name'], as_index=False).agg({'visitors':'mean'})

fig = px.scatter(p1, x='visitors', y='air_genre_name', color='wday')
fig.update_traces(marker={'size': 15})

fig.for_each_trace(lambda t: t.update(name=t.name.replace("0", "Monday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("1", "Tuesday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("2", "Wednesday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("3", "Thursday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("4", "Friday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("5", "Saturday")))
fig.for_each_trace(lambda t: t.update(name=t.name.replace("6", "Sunday")))

fig.update_layout(height=1000, width=800)
fig.show()

foo = air_visits.merge(air_store, on = 'air_store_id', how='left')
genres = list(foo.groupby('air_genre_name'))
subplot_titles = [genre[0] for genre in genres]

fig=go.Figure()

for i, l in enumerate(genres):
    df = l[1]
    fig.add_trace(go.Box(x=df['visitors'], showlegend=False, name=subplot_titles[i]))

fig.update_xaxes(title='Visitor Count (log)', type='log')
fig.update_layout(height=600, width=1000, title='Visitor Distribution by Genre (AIR)')
fig.show()

foo = 1
bar = 1
baz = 1
p1 = 1
p2 = 1
p3 = 1
df_tmp = 1

Findings:

• From the day of week plot, we can see that light blue and pink points (representing Saturday and Sunday) are typically higher on the average visitor count than the rest of the week. The genres where this is not true are Yakiniku/Izikaya/Dining Bar, all which have Friday and Saturdays as the higher visitor counts
• The genre with the largest gap in average visitors is "Karaoke", where the weekends dominate the rest of the week. "International cuisine" shows a similar trend but the gap is substantially smaller
• "Bars/cocktailis" notably lower overall in terms of popularity with "Asian" having overall the highest average visitors
• The box plots confirm a few things we saw from the prior graph: "Asian" restaurants have the highest meadian at 38 and the bulk of data being above 25 visitors daily, "Karaoke" also once again shows up with the largest Q1 and Q3 spread showing the volitility and impact of the weekend visitors on a normally low visitor count week

We can take a deeper look in relation to the holiday data we previously saw, let's see the impact of holidays on average visitor data overall rather than by category for now:

foo = air_visits.rename(columns={'visit_date':'calendar_date'})
bar = foo.merge(holidays, on = 'calendar_date', how='left')
bar['holiday_flg'] = bar['holiday_flg'].astype(bool)

fig = px.box(bar, x='holiday_flg', y='visitors', color='holiday_flg')
fig.update_yaxes(type='log')
fig.show()

bar['wday'] = pd.to_datetime(bar['calendar_date']).dt.dayofweek
p1 = bar.groupby(['wday', 'holiday_flg'], as_index=False).agg({'visitors': 'mean'})
wdays= ['Mon', 'Tues', 'Wed', 'Thurs', 'Fri', 'Sat', 'Sun']

fig = px.scatter(p1, x='wday', y='visitors', color='holiday_flg')
fig.update_traces(marker={'size': 15})
fig.update_layout(height=600, width=600)
fig.layout.xaxis.tickvals = [i for i in range(len(wdays))]
fig.layout.xaxis.ticktext = wdays 
fig.show()

Based on the first plot, I don't see much of a shift in average visitors holiday vs. not holiday. To see this with a bit more clarity, the second plot shows the breakdown by day of week and we see that holidays that land on weekdays (Mon-Fri) have a higher average visitor count with the weekend showing a negative impact on Saturday and only a slight positive impact on Sunday.